Customized Measurement

In this example, we write a new AttackMetric to evaluate attackers and report average input length.

Write a New AttackMetric

1
2
3
4
5
class SentenceLength(OpenAttack.AttackMetric):
    NAME = "Input Length"

    def after_attack(self, input, adversarial_sample):
        return len(input["x"].split(" "))

We extend AttackMetric and override after_attack method to report the length of attacked sentence (assume that words are separated by white spaces).

NAME attribute indicates the name of the AttackMetric that will show out in the final result.

Apply AttackMetrics in AttackEval

1
2
3
attack_eval = OpenAttack.AttackEval(attacker, clsf, metrics=[
    SentenceLength()
])

AttackEval supports the metrics option to specify evaluation metrics.

There are some built-in metrics in OpenAttack:

  • OpenAttack.metric.Fluency()

  • OpenAttack.metric.GrammaticalErrors()

  • OpenAttack.metric.SemanticSimilarity()

  • OpenAttack.metric.EditDistance()

  • OpenAttack.metric.ModificationRate()

These metrics are common in most attack evaluations.

Complete Code

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
import OpenAttack
import datasets
class SentenceLength(OpenAttack.AttackMetric):
    NAME = "Input Length"
    def after_attack(self, input, adversarial_sample):
        return len(input["x"].split(" "))
def main():
    victim = OpenAttack.loadVictim("BERT.SST")
    def dataset_mapping(x):
        return {
            "x": x["sentence"],
            "y": 1 if x["label"] > 0.5 else 0,
        }
    dataset = datasets.load_dataset("sst", split="train[:20]").map(function=dataset_mapping)
    attacker = OpenAttack.attackers.GeneticAttacker()
    attack_eval = OpenAttack.AttackEval(attacker, victim, metrics=[
        SentenceLength()
    ])
    attack_eval.eval(dataset, visualize=True)

Run python examples/custom_metrics.py to see visualized results.